dfrgrgcccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc...
TRANSCRIPT
-
8/22/2019 dfrgrgccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc
http:///reader/full/dfrgrgccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc 1/2
Power-Aware Routing and Wavelength Assignment in Optical Networks
Yong Wu, Luca Chiaraviglio, Marco Mellia, Fabio Neri
Politecnico di Torino, Corso Duca Degli Abruzzi 24, 10129, Torino, Italy B [email protected]
Abstract We introduce the Power-Aware RWA problem, whose goal is to accommodate lightpaths in wavelength
routing networks minimizing the power consumption. Formulation, algorithms, and results are presented, showing
that significant power savings are possible.
Introduction
Wavelength Routing (WR) networks offer the flexibility
of designing a logical topology, comprising lightpath
requests, over a physical topology, comprising OXCs
and links with many fibers each. The Routing and
Wavelength Assignment (RWA) problem is well known
in the literature [1]. Its goal is to assign a route and
a suitable wavelength in the physical topology for eachlightpath of the logical topology. Traditionally, the goal
of the RWA problem is to minimize the load (e.g., num-
ber of wavelengths) on available resources, in order
to maximize the probability of accommodating possible
new lightpath requests. However, this leads in general
to a waste in the power required to keep up and run-
ning both OXCs and optical amplifiers along fiber links.
Given the large number of these devices (thousands)
and their power footprint (up to tens of kW), we propose
to target the minimization of power consumption when
solving the RWA problem, by making maximum usage
of powered-on devices, e.g., by reusing the same fiber
along the same path as much as possible, in contrast to
spreading lightpaths on available fibers and paths. We
name this problem Power-Aware RWA (PA-RWA) prob-
lem. In this paper, we give a formulation of the problem,
propose heuristics to solve it, and present simulation
results showing that a large amount of power can be
saved in WR networks, reducing up to a factor of 5 the
energy (and the cost) needed to operate a WR network.
Problem Formulation
The PA-RWA problem can be defined using an integer
linear programming (ILP) formulation. Let sd denote
the number of lightpath requests from source s to des-tination d, and sdw the number of lightpaths from s to
d on wavelength w: sd =
wsdw. Let f
sdwijk {0, 1}
denote the number of lightpaths from s to d on fiber k
of link (i, j) using wavelength w, and fsdwij =
kfsdwijk .
On link (i, j), let Kij be the number of fibers, and Fijkthe number of wavelengths available on fiber k. Let aijk
be the number of amplifiers on fiber k of link (i, j), and
xijk {0, 1} be binary variables equal to 1 if fiber k
on link (i, j) is used to route a lightpath. Similarly, let
yi {0, 1} be binary variables equal to 1 if OXC i is
used. Finally, let PA and PO be the power consump-
tions of one amplifier and one OXC, respectively. We
do not consider the possibility of powering off individ-
This work was partially supported by the BONE research project.
ual transceivers or subsystems in OXCs. The notation
above leads to the following ILP formulation of the PA-
RWA problem:
min Ptot; Ptot = PAi,j,k
aijkxijk + POi
yi (1)
s.t.s,d
fsdwijk 1 w,i,j ,k (2)
s,d,w
fsdwijk M xijk i,j ,k (3)
j,k
xijk +j,k
xjik Myi i (4)
k
fsdwijk = f
sdwij s ,d,w,i,j (5)
i
fsdwij f
sdwji
=
sdw, j = s
sdw, j = d s ,d,w,j
0, j = s, d
(6)
s,d,w
fsdwijk Fijk i,j ,k;
k
xijk Kij i, j (7)
Eq. 1 is the utility function. Eq. 2 constraints a wave-
length on a fiber to be assigned to at most one light-
path; Eq. 3 imposes that a fiber is used if at least one
wavelength is assigned (M, and later M, are sufficient
large constants); Eq. 4 state that OXC i has to be pow-
ered on if any of its fibers is used; Eq. 5 counts the
number of fibers; Eq. 6 are the classical routing con-
straints, assuming no wavelength conversion function-
ality; finally Eqs. 7 limit the number of wavelengths on
each fiber, and the number of fibers in each link.
Algorithms
Following a well-established approach, the problem is
divided in two parts: routing and fiber&wavelength as-
signment. Considering the routing problem, we imple-
mented three heuristics: Least-Cost Path (LCP) [1],
Most-Used Path (MUP) and Ordered-Lightpath Most-
Used Path (OLMUP). All algorithms start with a network
where yi = 0 and xijk = 0 i,j ,k; the initial cost for
each link is cij = aijPA + PO (assuming for simplicity
that aijk = aij k). The main steps of each algorithm
are described below.
LCP: for each request, compute the shortest path P
using cij as static routing costs.
MUP: for each request, compute the shortest path up-
dating the routing costs of links used to route the cur-rent lightpath request. Using a pseudo-code notation,
for each lightpath sd
mailto:[email protected]:[email protected]:[email protected] -
8/22/2019 dfrgrgccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc
http:///reader/full/dfrgrgccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc 2/2
0
50
100
150
200
250
300
350
5 10 15 20 25 30
PTOT[kW]
N
LCP-FF
MUP-FF
OLMUP-FF
LB
Fig. 1: Power Consumption vs number of nodes randomphysical topology
compute the shortest path Psd using costs cij
cij = 0 (i, j) PsdOLMUP: at each iteration, select the lightpath to be
routed as the one that minimizes the incremental cost
(Lightpath Selection LS phase). Then route it us-
ing the MUPalgorithm (Routing Update RU phase).During LS, find the lightpath that has the minimum in-
cremental cost: for each lightpath sd not yet assigned
compute the shortest path Psd using costs cij
compute the incremental cost CP =
ijPsdcij
if CP < Cmin then Cmin = CP; s = s; d = d;
P = Psd; = sd
During RU, route on P using the MUPalgorithm.
Considering the fiber&wavelength assignment prob-
lem, we use in all cases a simple First-Fit strategy
(FF) which has been proved to be one of the most ef-
fective algorithms [1]: for each lightpath, considering
first powered-on fibers, wavelengths are sequentiallyscanned for availability on the whole path. If no wave-
length is found, a new fiber is powered on in each link.
In case no assignment is possible, a new path is com-
puted, forbidding links in the previously chosen path.
Finally, for all used fibers and OXCs, set xijk = 1 and
yi = 1 and compute Ptot as in Eq. 1.
To compare results, we define a rough lower bound
(LB) by considering that at least all OXCs which are
source or destination of a lightpath request must be
powered on, and that at least the cheapest fiber must
be used to exit from s or to arrive to d for each lightpath.
Performance Evaluation
Several scenarios were studied in [2]. We report here
results considering physical topologies which are either
bidirectional rings, or generic meshes (generated con-
sidering a probability 0.5 of having a link between any
two OXCs). We assume Kij = 10 and Fijk = 128,
and PA = PO = 1kW. The number of amplifiers on
each fiber aijk is instead a discrete random variable
uniformly distributed in [0, 10].
We first consider a set of randomly chosen light-
paths, so that there is a lightpath between any two
nodes with probability 0.5. Fig. 1 reports the total power
consumption required to route all lightpath requestsversus the number of OXCs N in the physical topology.
Results are obtained averaging over 10 independent
0
200
400
600
800
1000
1200
1400
1600
1800
5 10 15 20 25 30
PTOT[kW]
N
LCP-FF
MUP-FF
OLMUP-FF
LB
Fig. 2: Power Consumption vs number of nodes bidirec-tional ring physical topology
02468
10
500 550 600 650 700
Kon
ij
(i,j)ID
LCP-FFMUP-FF
OLMUP-FF
02468
10
0 10 20 30 40 50 60
Kon
ij
(i,j)IDFig. 3: Number of used fibers vs link ID random topology onthe left, and bidirectional ring on the right. N = 32.
runs. Using LCP-FF, which is almost power-unaware,
the power consumption of the network grows as 10N,
since no spatial reuse is enforced. Considering on the
contrary both the MUP-FF and OLMUP-FF heuristics,
much better results are obtained, due to the power-
aware routings, with the OLMUP-FF being very close
to the lower bound. The energy saving can be very sig-
nificant even for small networks, e.g., for N = 24 nodes
the power consumption is approximately reduced by a
factor of 5, i.e., around 200kW of saving.
Considering a bidirectional ring physical topology,Fig. 2 details PTOT versus N: also in this case
OLMUP-FF shows the best results. Notice that the total
power consumption is much larger in this scenario, and
the lower bound is not very tight. This is due to the reg-
ularity of the physical topology, in which path lengths
are larger [O(N) versus O(ln(N))], and to less alter-
natives in path selection. Still power saving achieved
by OLMUP-FF is larger than 400kW if N = 24.
Fig. 3 reports the number Konij of fibers powered on
for each link (omitting links where Kon = 0). It shows
that LCP-FF always uses more fibers (therefore more
power) than the other two heuristics which on the con-trary try to reuse as much as possible already used
links. This holds true for both physical topologies. In
case of the bidirectional physical topology, a larger per-
centage of links are required as expected.
While the considered scenario, with large numbers
of fibers on each link and of wavelengths per fiber com-
pared to the number of lightpaths, eases the reduction
of devices that must be powered on, our results still
bring evidence to the fact that new design criteria aimed
at power saving can be very effective.
References
1 H. Zang at al., A Review of RWA Approaches for WR Net-works, SPIE Optical Networks Mag, 2000
2 Y.Wu, Ms thesis; http://www.tlc.polito.it/wu.pdf.
http://www.tlc.polito.it/wu.pdfhttp://www.tlc.polito.it/wu.pdfhttp://www.tlc.polito.it/wu.pdf